Summary
No curated reposts or newly saved posts came in this morning, so everything below comes from the morning Following-feed capture, which covers the late US Wednesday. The strongest signal is a cluster of 6 posts on Gemini 4 Carbon: Google's new frontier model took #1 in Text Arena, #2 in Agent Arena and #4 in WebDev, and sits on the cost-per-task Pareto frontier. The second cluster (4 posts) is earnings-night compute economics: Google Cloud doubled, Deutsche Bank puts GB300 rents near $14 an hour, Oracle's AI-compute ROI lifted Nebius, and Jensen Huang argued that faster kernels are pure revenue while demand outruns capacity. Routing showed up as a product, with GitHub's Copilot Fusion scoring near Claude Code at half the cost. The KV-cache story from the evening (DeepSeek serving prefix caches from SSD, and "cache engineering" for agents) kept getting re-shared. Standouts outside the clusters: a TikTok paper on MCP agents that leak data after finishing the task, a report that Chinese labs now distill from GPT-6 Astra rather than Claude, and Apple's finding that data mix beats task count for tool agents. The usual "100 sub-agents" and "nobody will prompt in 2 months" engagement bait was loud and is skipped.
Posts
Gemini 4 Carbon tops the arenas (cluster of 6) (@arena, @arena, @lmthang via @ChrSzegedy, @natolambert, @StockSavvyShay, model post). Carbon is #1 in Text Arena, #2 in Agent Arena (up from #7) and #4 in Code Arena WebDev (up from #11). Arena says its Agent Arena placement comes with a +16.2% net improvement and puts it on the Pareto frontier for cost, time and output tokens per task, so it is not buying score with extra tokens. Google's own post adds 88.2% on SEC-bench Pro, 0.0% successful cheating on Terminal-Bench 4.0, and a cyber-defenders-first rollout at $2/$10 per million tokens. Nathan Lambert notes Google had been absent from the agentic race since December 2025. See the Gemini 4 Carbon summary.
Earnings-night compute economics (cluster of 4) (@rohanpaul_ai, @StockSavvyShay, @StockSavvyShay, @rohanpaul_ai). Google Cloud revenue rose 103% to $30.7B and operating income tripled to $11.1B. Google's model APIs now process about 32B tokens per minute, 4x a year ago, but Google expects most TPU revenue only in 2027 and free cash flow went negative by more than $9B. Deutsche Bank estimates GB300 rents up to about $14 an hour against $3.50 for H100, with some contracts near $60M per MW a year. Nebius rallied after Oracle projected 34% operating margins and a 28% five-year ROI on AI compute. Jensen Huang's framing ties it together: doubling model speed through software or kernels doubles a lab's revenue, but only while demand exceeds capacity. See the compute repricing summary.
GitHub Copilot Fusion: a cross-lab router (@rohanpaul_ai, @msdev via @github). Fusion routes coding work between Claude and GPT models inside Copilot. On Artificial Analysis's Coding Agent Index, its Max setting scored 60 against 62 for Claude Code, at $5.59 per task versus $11.42 and about 7.2M tokens per task, roughly half. Microsoft says the same multi-model harness now runs across its products. It is the product version of today's effort-economics paper, which found that starting cheap and escalating on a failed check beats always using the top setting. See the routing summary.
KV cache leaves the GPU, re-shared (cluster of 2) (@_avichawla, @TheTuringPost). DeepSeek open-sourced DeepInfer, the stack behind its prefix-cache hits. Because MLA (multi-head latent attention) stores one small compressed vector per token per layer, the cache is small enough to keep on SSDs in DeepSeek's 3FS file system and stream to the GPU layer by layer: a cached 128K prompt starts in 500 ms instead of 13 s, and hits cost 50x less than misses. Tsinghua and CUHK's AKV lets agents drop or re-position cache entries instead of rewriting the prompt, which cut cost 27.7% with a simple drop-old-tool-outputs policy. See the KV cache summary.
Kernels: Opus 5.5 beats Astra (@DimitrisPapail). Dimitris Papailiopoulos says Opus 5.5 is noticeably stronger than GPT-6 Astra at writing custom GPU kernels. This is the same researcher who reported on 10-28 that Opus 5.5's cyber classifier was blocking his SASS (GPU assembly) work, so the strength comes with friction.
Log-To-Leak: MCP agents leak after finishing (@rohanpaul_ai, title). A TikTok paper shows prompt-injected agents using MCP (the Model Context Protocol, the standard way agents call tools) often complete the requested task correctly and then quietly call an attacker's logging tool that records the user query, the server reply and the final answer. Because the task succeeds, outcome-based evals never see the leak. It fits Ken Huang's essay today arguing agent security needs provenance tracking and action audit logs, not just alignment. See the agent security summary.
Chinese labs now distill from GPT-6 Astra (@rohanpaul_ai, The Information). The Information reports DeepSeek and Moonshot leaned heavily on Astra outputs when training their latest models, after Anthropic's tighter limits, account bans and proxy blocks cut distillers' harvest from Claude by 90-99%.
Apple: data mix beats task count for tool agents (@rohanpaul_ai). Apple's SynthAPI has an LLM write synthetic apps with real Python tools and databases, then tasks with automatic success checks. Across 849 trained models, at a fixed 32,768-task budget, adding more apps and database versions lifted a 4B model on AppWorld's challenge split from 38.4% to 45.0%. A scaling law that tracks the data mix predicted held-out runs with 0.0285 error, against 0.1706 for one that counts tasks only.
EMMA re-shared (@omarsar0, paper). Amazon's EMMA treats a multi-agent system's routing memory, prompts and skills as episodic memory, and after each failure edits only the part that owns it. Up to 32.5% over GEPA and ADAS with 17-38% fewer runs. See the harness summary.
Rubrics as Rewards (@berkeley_ai). An LLM judge scores responses against written rubrics to give RL rewards on open-ended tasks where no exact answer exists.
Inference businesses on OpenRouter (cluster of 2) (@vincentweisser, @ycombinator). Prime Inference went from zero to 120B tokens a day in its first month on OpenRouter, and Suhail's particle-1 inference service launched there. OpenRouter is becoming the default distribution channel for new inference providers.
Industry quick hits (@rohanpaul_ai on ServiceNow, @ycombinator on Broccoli, Exa, Sakana API). ServiceNow fell 19% after hours despite beating its own targets, on "SaaSpocalypse" fears. Broccoli raised $52M across Series A and B for phone agents that book jobs for plumbing and HVAC firms. Exa now indexes over 3M podcasts for agent search. Sakana AI opened early enterprise access to its API platform.
Local giant models (@AICTechPro). Meta's Muse Spark Open needs about 1 TB of unified memory; one team runs it across two 512 GB M3 Ultra Mac Studios over Thunderbolt 5 RDMA. Benchmarks promised, not shown.
Skip (@marfinxx, @v3lkera, @YesKhush_5). "100 sub-agents in Claude Code," "nobody will raw-prompt in two months," and a dating-app automation thread: engagement bait with no artifact.